Re: ixgbe zero-copy performance

On Wed, 10 Jul 2024 at 12:07, Srivats P <[email protected]> wrote:
>
> > > What is the expected performance for AF_XDP txpnly in zero-copy and copy modes?
> > >
> > > With Kernel 6.5.0 and the same ixgbe driver, this is what we see -
> > >
> > > ZC mode: 4.3Mpps
> > > Copy mode: 3.3Mpps
> > >
> > > This doesn't seem right. Shouldn't the zero copy performance be MUCH higher?
> >
> > Zero-copy performance should be line rate for the ixgbe card, so
> > somewhere around 15Mpps. SKB mode seems in the correct ballpark. Try
> > pinning the app to a core the driver does not run on, or use busy-poll
> > mode "-B". If you are running on a NUMA system, make sure you are
> > running both driver and app on the NUMA node you have plugged your NIC
> > into.
>
> I had forced zero-copy mode (-z) and copy mode (-c) to get the above
> results. So the xsk creation would have failed if it were not zero
> copy mode.

Just to avoid any confusion: copy mode = skb mode for Tx. The 3Mpps
seems about correct for that mode.

> This is not a NUMA system - just one CPU with 8 cores, HT
> disabled. Nothing else taking up CPU on the system, so I don't think
> the app and the softirq would have been sharing a core.

I would ask them so that this is not the case. It happens.

> Unfortunately, this is at a customer - so I have asked them to try
> taskset or use -B. Will update when I hear back from them.

Sounds good.

> Thanks Magnus!
>
> Srivats